Readme.md

Scalable ML Model Deployment

Atchuth Naveen Chilaparasetti

• LinkedIn | Website

System Design

For serving from Docker

docker-compose up

For serving from kubernetes

Setup MiniKube

minikube addons enable metrics-server

minikube start --extra-config kubelet.EnableCustomMetrics=true

Deploy Kubernetes containers

kubectl apply -f k8s

To start the service:

minikube service web-load-balancer-service

Testing Environment

For Frontend Test,

Server Url : http://127.0.0.1:8080/home

Test Url : https://tensorflow.org/images/blogs/serving/cat.jpg

Running Kube Tests

Set up Locust for Stress Testing

From the directory containing locustfile.py, run

locust

Testing UI on http://127.0.0.1:8089/

Tests

Performed tests using two model servers that take performs image calculation and a common web server that routes requests from the front end. Further directed requests to only one of the model servers and limited its max cpu utilization to 50% whereas the web server container can max cpu utilization of 90%. So, with more load, Kubernetes prefers to scale more instances of model servers.

With No Auto Scaling - Single instance of each

• Response rate - The response rate increases steadily and at 75 users, the system no longer can support.

• Resource Usage - At 75 users, the system crashes continuously and the web server goes down. But Kubernetes makes sure to restart the container as soon as it detects a failure.

With Auto Scaling Enabled using Horizontal Pod scaler

• Response rate

• Resource Usage at the start

• Resource Usage at the time of scaling - Kubernetes automatically spawns two more instances for model server 1 that is in use whereas the other model server is idle. Because the web server can have resource usage up to 90%, Kubernetes adjusted the resources between the model and web servers appropriately.

• Resource Usage during downscaling - When no requests are being sent, Kubernetes automatically performs downscaling

• Resource Usage from minikube dashboard

minikube dashboard

• Resource usage at full user load

Google Cloud

As all the tests are performed locally, there is a resource constraint to perform tests on a full scale. So, we migrated the code Google cloud and performed similar tests and the results are presented. For commands to set up gcloud Kubernetes engine and the test results, refer commands , report.